Beyond Additivity: Sparse Isotonic Shapley Regression toward Nonlinear Explainability